Python tensorflow.python.ops.control_flow_ops.group() Examples

def testRestoredModelPerformance(self):
    checkpoint_path = os.path.join(self.get_temp_dir(), 'model.ckpt')
    log_dir = os.path.join(self.get_temp_dir(), 'log_dir1/')

    # First, save out the current model to a checkpoint:
    init_op = control_flow_ops.group(variables.global_variables_initializer(),
                                     variables.local_variables_initializer())
    saver = saver_lib.Saver(write_version=saver_pb2.SaverDef.V1)
    with self.test_session() as sess:
      sess.run(init_op)
      saver.save(sess, checkpoint_path)

    # Next, determine the metric to evaluate:
    value_op, update_op = metric_ops.streaming_accuracy(self._predictions,
                                                        self._labels)

    # Run the evaluation and verify the results:
    accuracy_value = evaluation.evaluate_once(
        '', checkpoint_path, log_dir, eval_op=update_op, final_op=value_op)
    self.assertAlmostEqual(accuracy_value, self._expected_accuracy) 

def _finish(self, update_ops, name_scope):
        # Update the power accumulators.
        with ops.control_dependencies(update_ops):
            with ops.colocate_with(self._iterations):
                update_beta1 = self._beta1_power.assign(
                    self._beta1_power * self._beta1_t,
                    use_locking=self._use_locking)
                update_beta2 = self._beta2_power.assign(
                    self._beta2_power * self._beta2_t,
                    use_locking=self._use_locking)
                t = self._iterations + 1.
                update_iterations = self._iterations.assign(t, use_locking=self._use_locking)
                momentum_cache_power = self._get_momentum_cache(self._schedule_decay_t, t)
                momentum_cache_t = self._beta1_t * (1. - 0.5 * momentum_cache_power)
                update_m_schedule = self._m_schedule.assign(
                    self._m_schedule * momentum_cache_t,
                    use_locking=self._use_locking)
        return control_flow_ops.group(
            *update_ops + [update_beta1, update_beta2] + [update_iterations, update_m_schedule],
            name=name_scope) 

def _apply_dense(self, grad, var):
        lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
        beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
        beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)

        # the following equations given in [1]
        # m_t = beta1 * m + (1 - beta1) * g_t
        m = self.get_slot(var, "m")
        m_t = state_ops.assign(m, beta1_t * m + (1. - beta1_t) * grad, use_locking=self._use_locking)

        # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
        v = self.get_slot(var, "v")
        v_t = state_ops.assign(v, beta2_t * v + (1. - beta2_t) * tf.square(grad), use_locking=self._use_locking)
        v_prime = self.get_slot(var, "v_prime")
        v_t_prime = state_ops.assign(v_prime, tf.maximum(v_prime, v_t))

        var_update = state_ops.assign_sub(var,
                                          lr_t * m_t / (tf.sqrt(v_t_prime) + epsilon_t),
                                          use_locking=self._use_locking)

        return control_flow_ops.group(*[var_update, m_t, v_t, v_t_prime])

    # keras Nadam update rule 

def _init_local_init_op(self, local_init_op=USE_DEFAULT):
    """Initializes local_init_op.

    Args:
      local_init_op: `Operation` run for every new supervisor instance. If set
      to USE_DEFAULT, use the first op from the GraphKeys.LOCAL_INIT_OP
      collection. If the collection is empty, create an op that initializes
      all local variables and all tables.
    """
    if local_init_op is Supervisor.USE_DEFAULT:
      local_init_op = self._get_first_op_from_collection(
          ops.GraphKeys.LOCAL_INIT_OP)
      if local_init_op is None:
        op_list = [
            variables.local_variables_initializer(),
            lookup_ops.tables_initializer()
        ]
        if op_list:
          local_init_op = control_flow_ops.group(*op_list)
          ops.add_to_collection(ops.GraphKeys.LOCAL_INIT_OP, local_init_op)
    self._local_init_op = local_init_op 

def _apply_sparse(self, grad, var):
        lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
        beta_t = math_ops.cast(self._beta_t, var.dtype.base_dtype)
        alpha_t = math_ops.cast(self._alpha_t, var.dtype.base_dtype)

        eps = 1e-7  # cap for moving average

        m = self.get_slot(var, "m")
        m_slice = tf.gather(m, grad.indices)
        m_t = state_ops.scatter_update(m, grad.indices,
                                       tf.maximum(beta_t * m_slice + eps, tf.abs(grad.values)))
        m_t_slice = tf.gather(m_t, grad.indices)

        var_update = state_ops.scatter_sub(var, grad.indices,
                                           lr_t * grad.values * (
                                                   1.0 + alpha_t * tf.sign(grad.values) * tf.sign(m_t_slice)))

        # Create an op that groups multiple operations
        # When this op finishes, all ops in input have finished
        return control_flow_ops.group(*[var_update, m_t]) 

def main_op():
  """Returns a main op to init variables and tables.

  Returns the main op including the group of ops that initializes all
  variables, initializes local variables and initialize all tables.

  Returns:
    The set of ops to be run as part of the main op upon the load operation.
  """
  init = variables.global_variables_initializer()
  init_local = variables.local_variables_initializer()
  init_tables = lookup_ops.tables_initializer()
  return control_flow_ops.group(init, init_local, init_tables)


# TODO(sukritiramesh): Integrate with Saver for complete restore functionality. 

def main_op_with_restore(restore_op_name):
  """Returns a main op to init variables, tables and restore the graph.

  Returns the main op including the group of ops that initializes all
  variables, initialize local variables, initialize all tables and the restore
  op name.

  Args:
    restore_op_name: Name of the op to use to restore the graph.

  Returns:
    The set of ops to be run as part of the main op upon the load operation.
  """
  with ops.control_dependencies([main_op()]):
    main_op_with_restore = control_flow_ops.group(restore_op_name)
  return main_op_with_restore 

def _MultiDeviceAddN(tensor_list):
  """Adds tensors from potentially multiple devices."""
  # Basic function structure comes from control_flow_ops.group().
  # Sort tensors according to their devices.
  tensors_on_device = collections.defaultdict(lambda: [])
  for tensor in tensor_list:
    tensors_on_device[tensor.device].append(tensor)

  # For each device, add the tensors on that device first.
  # Then gather the partial sums from multiple devices.
  # TODO(sjhwang): Create hierarchical aggregation tree as pbar's suggestion.
  # E.g., aggregate per GPU, then per task, and so on.
  summands = []

  def DeviceKey(dev):
    return "" if dev is None else dev

  for dev in sorted(six.iterkeys(tensors_on_device), key=DeviceKey):
    tensors = tensors_on_device[dev]
    with ops.colocate_with(tensors[0].op, ignore_existing=True):
      summands.append(math_ops.add_n(tensors))

  return math_ops.add_n(summands) 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def _finish(self, update_ops, name_scope):
    """Do what is needed to finish the update.

    This is called with the `name_scope` using the "name" that
    users have chosen for the application of gradients.

    Args:
      update_ops: List of `Operation` objects to update variables.  This list
        contains the values returned by the `_apply_dense()` and
        `_apply_sparse()` calls.
      name_scope: String.  Name to use for the returned operation.

    Returns:
      The operation to apply updates.
    """
    return control_flow_ops.group(*update_ops, name=name_scope)

  # --------------
  # Utility methods for subclasses.
  # -------------- 

def _init_local_init_op(self, local_init_op=USE_DEFAULT):
    """Initializes local_init_op.

    Args:
      local_init_op: `Operation` run for every new supervisor instance. If set
      to USE_DEFAULT, use the first op from the GraphKeys.LOCAL_INIT_OP
      collection. If the collection is empty, create an op that initializes
      all local variables and all tables.
    """
    if local_init_op is Supervisor.USE_DEFAULT:
      local_init_op = self._get_first_op_from_collection(
          ops.GraphKeys.LOCAL_INIT_OP)
      if local_init_op is None:
        op_list = [
            variables.local_variables_initializer(),
            lookup_ops.tables_initializer()
        ]
        if op_list:
          local_init_op = control_flow_ops.group(*op_list)
          ops.add_to_collection(ops.GraphKeys.LOCAL_INIT_OP, local_init_op)
    self._local_init_op = local_init_op 

def scatter_update(cls, factor, indices, values, sharding_func, name=None):
    """Helper function for doing sharded scatter update."""
    assert isinstance(factor, list)
    if len(factor) == 1:
      with ops.colocate_with(factor[0]):
        # TODO(agarwal): assign instead of scatter update for full batch update.
        return state_ops.scatter_update(factor[0], indices, values,
                                        name=name).op
    else:
      num_shards = len(factor)
      assignments, new_ids = sharding_func(indices)
      assert assignments is not None
      assignments = math_ops.cast(assignments, dtypes.int32)
      sharded_ids = data_flow_ops.dynamic_partition(new_ids, assignments,
                                                    num_shards)
      sharded_values = data_flow_ops.dynamic_partition(values, assignments,
                                                       num_shards)
      updates = []
      for i in xrange(num_shards):
        updates.append(state_ops.scatter_update(factor[i], sharded_ids[i],
                                                sharded_values[i]))
      return control_flow_ops.group(*updates, name=name) 

def __init__(self, inputs, outputs, updates=None):
    updates = updates or []
    if not isinstance(inputs, (list, tuple)):
      raise TypeError('`inputs` to a TensorFlow backend function '
                      'should be a list or tuple.')
    if not isinstance(outputs, (list, tuple)):
      raise TypeError('`outputs` of a TensorFlow backend function '
                      'should be a list or tuple.')
    if not isinstance(updates, (list, tuple)):
      raise TypeError('`updates` in a TensorFlow backend function '
                      'should be a list or tuple.')
    self.inputs = list(inputs)
    self.outputs = list(outputs)
    with ops.control_dependencies(self.outputs):
      updates_ops = []
      for update in updates:
        if isinstance(update, tuple):
          p, new_p = update
          updates_ops.append(state_ops.assign(p, new_p))
        else:
          # assumed already an op
          updates_ops.append(update)
      self.updates_op = control_flow_ops.group(*updates_ops) 

def _export_graph(graph, saver, checkpoint_path, export_dir,
                  default_graph_signature, named_graph_signatures,
                  exports_to_keep):
  """Exports graph via session_bundle, by creating a Session."""
  with graph.as_default():
    with tf_session.Session('') as session:
      variables.local_variables_initializer()
      lookup_ops.tables_initializer()
      saver.restore(session, checkpoint_path)

      export = exporter.Exporter(saver)
      export.init(
          init_op=control_flow_ops.group(
              variables.local_variables_initializer(),
              lookup_ops.tables_initializer()),
          default_graph_signature=default_graph_signature,
          named_graph_signatures=named_graph_signatures,
          assets_collection=ops.get_collection(ops.GraphKeys.ASSET_FILEPATHS))
      return export.export(export_dir, contrib_variables.get_global_step(),
                           session, exports_to_keep=exports_to_keep) 

def _apply_sparse(self, grad, var):
        lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
        alpha_t = math_ops.cast(self._alpha_t, var.dtype.base_dtype)
        beta_t = math_ops.cast(self._beta_t, var.dtype.base_dtype)

        eps = 1e-7  # cap for moving average

        m = self.get_slot(var, "m")
        m_slice = tf.gather(m, grad.indices)
        m_t = state_ops.scatter_update(m, grad.indices,
                                       tf.maximum(beta_t * m_slice + eps, tf.abs(grad.values)))
        m_t_slice = tf.gather(m_t, grad.indices)

        var_update = state_ops.scatter_sub(var, grad.indices, lr_t * grad.values * tf.exp(
            tf.log(alpha_t) * tf.sign(grad.values) * tf.sign(m_t_slice)))  # Update 'ref' by subtracting 'value
        # Create an op that groups multiple operations.
        # When this op finishes, all ops in input have finished
        return control_flow_ops.group(*[var_update, m_t]) 

def _init_local_init_op(self, local_init_op=USE_DEFAULT):
    """Initializes local_init_op.

    Args:
      local_init_op: `Operation` run for every new supervisor instance. If set
      to USE_DEFAULT, use the first op from the GraphKeys.LOCAL_INIT_OP
      collection. If the collection is empty, create an op that initializes
      all local variables and all tables.
    """
    if local_init_op is Supervisor.USE_DEFAULT:
      local_init_op = self._get_first_op_from_collection(
          ops.GraphKeys.LOCAL_INIT_OP)
      if local_init_op is None:
        op_list = [variables.local_variables_initializer(),
                   data_flow_ops.tables_initializer()]
        if op_list:
          local_init_op = control_flow_ops.group(*op_list)
          ops.add_to_collection(ops.GraphKeys.LOCAL_INIT_OP, local_init_op)
    self._local_init_op = local_init_op 

def _finish(self, update_ops, name_scope):
    """Do what is needed to finish the update.

    This is called with the `name_scope` using the "name" that
    users have chosen for the application of gradients.

    Args:
      update_ops: List of `Operation` objects to update variables.  This list
        contains the values returned by the `_apply_dense()` and
        `_apply_sparse()` calls.
      name_scope: String.  Name to use for the returned operation.

    Returns:
      The operation to apply updates.
    """
    return control_flow_ops.group(*update_ops, name=name_scope)

  # --------------
  # Utility methods for subclasses.
  # -------------- 

def add_check_numerics_ops():
  """Connect a `check_numerics` to every floating point tensor.

  `check_numerics` operations themselves are added for each `half`, `float`,
  or `double` tensor in the graph. For all ops in the graph, the
  `check_numerics` op for all of its (`half`, `float`, or `double`) inputs
  is guaranteed to run before the `check_numerics` op on any of its outputs.

  Returns:
    A `group` op depending on all `check_numerics` ops added.
  """
  check_op = []
  # This code relies on the ordering of ops in get_operations().
  # The producer of a tensor always comes before that tensor's consumer in
  # this list. This is true because get_operations() returns ops in the order
  # added, and an op can only be added after its inputs are added.
  for op in ops.get_default_graph().get_operations():
    for output in op.outputs:
      if output.dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
        message = op.name + ":" + str(output.value_index)
        with ops.control_dependencies(check_op):
          check_op = [array_ops.check_numerics(output, message=message)]
  return control_flow_ops.group(*check_op) 

def variables_initializer(var_list, name="init"):
  """Returns an Op that initializes a list of variables.

  After you launch the graph in a session, you can run the returned Op to
  initialize all the variables in `var_list`. This Op runs all the
  initializers of the variables in `var_list` in parallel.

  Calling `initialize_variables()` is equivalent to passing the list of
  initializers to `Group()`.

  If `var_list` is empty, however, the function still returns an Op that can
  be run. That Op just has no effect.

  Args:
    var_list: List of `Variable` objects to initialize.
    name: Optional name for the returned operation.

  Returns:
    An Op that run the initializers of all the specified variables.
  """
  if var_list:
    return control_flow_ops.group(*[v.initializer for v in var_list], name=name)
  return control_flow_ops.no_op(name=name) 

def _MultiDeviceAddN(tensor_list):
  """Adds tensors from potentially multiple devices."""
  # Basic function structure comes from control_flow_ops.group().
  # Sort tensors according to their devices.
  tensors_on_device = collections.defaultdict(lambda: [])
  for tensor in tensor_list:
    tensors_on_device[tensor.device].append(tensor)

  # For each device, add the tensors on that device first.
  # Then gather the partial sums from multiple devices.
  # TODO(sjhwang): Create hierarchical aggregation tree as pbar's suggestion.
  # E.g., aggregate per GPU, then per task, and so on.
  summands = []

  def DeviceKey(dev):
    return "" if dev is None else dev

  for dev in sorted(six.iterkeys(tensors_on_device), key=DeviceKey):
    tensors = tensors_on_device[dev]
    with ops.colocate_with(tensors[0].op, ignore_existing=True):
      summands.append(math_ops.add_n(tensors))

  return math_ops.add_n(summands) 

def main_op_with_restore(restore_op_name):
  """Returns a main op to init variables, tables and restore the graph.

  Returns the main op including the group of ops that initializes all
  variables, initialize local variables, initialize all tables and the restore
  op name.

  Args:
    restore_op_name: Name of the op to use to restore the graph.

  Returns:
    The set of ops to be run as part of the main op upon the load operation.
  """
  with ops.control_dependencies([main_op()]):
    main_op_with_restore = control_flow_ops.group(restore_op_name)
  return main_op_with_restore 

def apply_gradients(self, grads_and_vars, global_step=None, name=None):
    with ops.name_scope(name, self._name) as name:
      update_op = self._opt.apply_gradients(
          grads_and_vars, global_step=global_step)
      clip_update_ops = []
      with ops.control_dependencies([update_op]):
        for grad, var in grads_and_vars:
          if grad is None or var not in self._vars_to_clip_dims:
            continue
          with ops.name_scope("clip_" + var.op.name):
            if isinstance(grad, ops.Tensor):
              clip_update_ops.append(self._clip_dense(var))
            else:
              clip_update_ops.append(self._clip_sparse(grad, var))

      # In case no var was clipped, still need to run the update_op.
      return control_flow_ops.group(*([update_op] + clip_update_ops), name=name) 

def close(self, cancel_pending_enqueues=False, name=None):
    """Closes the barrier and the FIFOQueue.

    This operation signals that no more segments of new sequences will be
    enqueued. New segments of already inserted sequences may still be enqueued
    and dequeued if there is a sufficient number filling a batch or
    allow_small_batch is true. Otherwise dequeue operations will fail
    immediately.

    Args:
      cancel_pending_enqueues: (Optional.) A boolean, defaulting to
        `False`. If `True`, all pending enqueues to the underlying queues will
        be cancelled, and completing already started sequences is not possible.
      name: Optional name for the op.

    Returns:
      The operation that closes the barrier and the FIFOQueue.
    """
    with ops.name_scope(name, "SQSSClose", [self._prefetch_op]) as name:
      barrier_close = self.barrier.close(cancel_pending_enqueues,
                                         "BarrierClose")
      fifo_queue_close = self._capacity_queue.close(cancel_pending_enqueues,
                                                    "FIFOClose")
      return control_flow_ops.group(barrier_close, fifo_queue_close, name=name) 

def _train_op(loss, labels, train_op_fn, centered_bias, batch_size, loss_fn,
              weights):
  """Returns op for the training step."""
  if centered_bias is not None:
    centered_bias_step = _centered_bias_step(
        centered_bias=centered_bias,
        batch_size=batch_size,
        labels=labels,
        loss_fn=loss_fn,
        weights=weights)
  else:
    centered_bias_step = None
  with ops.name_scope(None, "train_op", (loss, labels)):
    train_op = train_op_fn(loss)
    if centered_bias_step is not None:
      train_op = control_flow_ops.group(train_op, centered_bias_step)
    return train_op 

def _extract_metric_update_ops(self, eval_dict):
    """Separate update operations from metric value operations."""
    update_ops = []
    value_ops = {}
    for name, metric_ops in six.iteritems(eval_dict):
      if isinstance(metric_ops, (list, tuple)):
        if len(metric_ops) == 2:
          value_ops[name] = metric_ops[0]
          update_ops.append(metric_ops[1])
        else:
          logging.warning(
              'Ignoring metric {}. It returned a list|tuple with len {}, '
              'expected 2'.format(name, len(metric_ops)))
          value_ops[name] = metric_ops
      else:
        value_ops[name] = metric_ops

    if update_ops:
      update_ops = control_flow_ops.group(*update_ops)
    else:
      update_ops = None

    return update_ops, value_ops 

def insert(self, keys, values, name=None):
    self._check_keys(keys)
    num_shards = self._num_shards
    if num_shards == 1:
      return self._table_shards[0].insert(keys, values, name=name)

    shard_indices = self._shard_indices(keys)
    # TODO(andreasst): support 'keys' that are not vectors
    key_shards = data_flow_ops.dynamic_partition(keys, shard_indices,
                                                 num_shards)
    value_shards = data_flow_ops.dynamic_partition(values, shard_indices,
                                                   num_shards)
    return_values = [
        self._table_shards[i].insert(key_shards[i], value_shards[i], name=name)
        for i in range(num_shards)
    ]

    return control_flow_ops.group(*return_values) 

def _get_local_init_op():
  """Returns the local init ops to initialize tables and local variables."""
  local_init_op = _get_first_op_from_collection(
      ops.GraphKeys.LOCAL_INIT_OP)
  if local_init_op is None:
    op_list = [
        variables.local_variables_initializer(),
        lookup_ops.tables_initializer()
    ]
    if op_list:
      local_init_op = control_flow_ops.group(*op_list)
      ops.add_to_collection(ops.GraphKeys.LOCAL_INIT_OP, local_init_op)
  return local_init_op 

def _apply_sparse_shared(self, grad, var, indices, scatter_add):
        beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
        beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
        lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
        beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
        beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)

        lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))

        # m_t = beta1 * m + (1 - beta1) * g_t
        m = self.get_slot(var, "m")
        m_scaled_g_values = grad * (1 - beta1_t)
        m_t = state_ops.assign(m, m * beta1_t, use_locking=self._use_locking)
        with ops.control_dependencies([m_t]):
            m_t = scatter_add(m, indices, m_scaled_g_values)

        # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
        v = self.get_slot(var, "v")
        v_scaled_g_values = (grad * grad) * (1 - beta2_t)
        v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
        with ops.control_dependencies([v_t]):
            v_t = scatter_add(v, indices, v_scaled_g_values)

        # amsgrad
        vhat = self.get_slot(var, "vhat")
        vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
        v_sqrt = math_ops.sqrt(vhat_t)
        var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t),
                                          use_locking=self._use_locking)
        return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t]) 

def restore(self, restored_tensors, restored_shapes):
    weights = restored_tensors[:len(restored_tensors) // 2]
    biases = restored_tensors[len(restored_tensors) // 2:]
    params = self._canonical_to_params(weights, biases)
    if not isinstance(params, tuple):
      params = (params,)
    assign_ops = [
        state_ops.assign(variable, param, validate_shape=False)
        for variable, param in zip(self._variables, params)
    ]
    return control_flow_ops.group(*assign_ops) 

def _auc_hist_accumulate(hist_true, hist_false, nbins, collections):
  """Accumulate histograms in new variables."""
  with variable_scope.variable_scope(
      None, 'hist_accumulate', [hist_true, hist_false]):
    # Holds running total histogram of scores for records labeled True.
    hist_true_acc = variable_scope.get_variable(
        'hist_true_acc',
        shape=[nbins],
        dtype=hist_true.dtype,
        initializer=init_ops.zeros_initializer(),
        collections=collections,
        trainable=False)
    # Holds running total histogram of scores for records labeled False.
    hist_false_acc = variable_scope.get_variable(
        'hist_false_acc',
        shape=[nbins],
        dtype=hist_true.dtype,
        initializer=init_ops.zeros_initializer(),
        collections=collections,
        trainable=False)

    update_op = control_flow_ops.group(
        hist_true_acc.assign_add(hist_true),
        hist_false_acc.assign_add(hist_false),
        name='update_op')

    return hist_true_acc, hist_false_acc, update_op